JIT: relop simplification during redundant branch opts by AndyAyersMS · Pull Request #127181 · dotnet/runtime

AndyAyersMS · 2026-04-20T20:54:15Z

Suppose we have a dominating branch A (with predicate pA) that shares a successor with a dominated branch B (with predicate pB).

can optimize away the comparison done in A.

Here we extend that optimization to handle some cases where pB does not imply pA, by forming (AND pB pA) (in VN space) and seeing if we can simplify it to a relop over the same operands as pB, or to a constant.

If so, we can remove the comparison in done A but now also must modify the comparison done in B.

For example

if ((x >= 100) && (x <= 100)) S;

can be simplified to

if (x == 100) S;

and

if ((x >= 100) || (x <= 100)) S;

can be simplified to

S;

As part of this, teach VN how to simplify various combinations of AND/OR/NOT involving relops (there are many cases).

Incorporates some of the changes from #83859, but does not try and handle "ignorable" side effects.

Fixes #98227.

Suppose we have a dominating branch A (with predicate pA) that shares a successor with a dominated branch B (with predicate pB). can optimize away the comparison done in A. Here we extend that optimization to handle some cases where pB does not imply pA, by forming (AND pB pA) (in VN space) and seeing if we can simplify it to a relop over the same operands as pB, or to a constant. If so, we can remove the comparison in done A but now also must modify the comparison done in B. For example ```if ((x >= 100) && (x <= 100)) S;``` can be simplified to ```if (x == 100) S;``` and ```if ((x >= 100) || (x <= 100)) S;``` can be simplified to ```S;``` As part of this, teach VN how to simplify various combinations of AND/OR/NOT involving relops (there are many cases). Incorporates some of the changes from dotnet#83859, but does not try and handle "ignorable" side effects. Fixes dotnet#98227.

AndyAyersMS · 2026-04-20T20:54:45Z

@EgorBo PTAL
fyi @dotnet/jit-contrib

Modest number of diffs locally.

Copilot

Pull request overview

This PR extends redundant-branch optimization in the CoreCLR JIT by using value-number (VN) boolean simplification to eliminate a dominating comparison even when the dominated predicate doesn’t directly imply it, by simplifying (pB AND pA) into a single relop (or constant) and updating the dominated comparison accordingly.

Changes:

Add VN simplifications for NOT over boolean forms (double-negation, DeMorgan transforms, and relop negation).
Add VN AND/OR simplification tables to combine related relops with identical operands into a tighter relop or a constant.
Extend optRedundantDominatingBranch to attempt VN-based simplification of blockPathVN AND domPathVN and (when successful) rewrite the dominated relop.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 6 comments.

File	Description
src/coreclr/jit/valuenum.cpp	Adds VN boolean/relop simplifications (NOT handling and AND/OR relop-combination tables).
src/coreclr/jit/redundantbranchopts.cpp	Tries VN simplification when direct implication fails; rewrites dominated compare when simplification yields a new relop.

Copilot · 2026-04-20T21:05:45Z

+                if (andApp.m_args[0] == pathApp.m_args[0] && andApp.m_args[1] == pathApp.m_args[1])
+                {
+                    newRelop = (genTreeOps)andApp.m_func;
+                }


andApp.m_func is a VNFunc and may be one of the unsigned relop VNFuncs (e.g., VNF_LT_UN, VNF_GE_UN), which are not valid genTreeOps values. Casting it to genTreeOps and later calling OpName/SetOper can misbehave. Map unsigned VNFuncs to the corresponding GT_LT/GT_GE/... and update the compare node's unsigned flag instead of casting directly.

Copilot · 2026-04-20T21:05:45Z

+            ValueNum  andVN = vnStore->VNForFunc(TYP_INT, VNF_AND, blockPathVN, domPathVN);
+            VNFuncApp andApp;
+            VNFuncApp pathApp;
+            if (vnStore->IsVNRelop(andVN, &andApp) && vnStore->GetVNFunc(blockPathVN, &pathApp))
+            {


The simplification path only handles the case where andVN simplifies to a relop. VNForFunc(AND, ...) can also simplify to a constant 0/1; if that happens, IsVNRelop(andVN, ...) will be false and you miss the optimization. Consider explicitly handling the constant case (e.g., by bashing/removing the dominated compare when the path condition becomes always-true/always-false).

Copilot · 2026-04-20T21:05:46Z

+                // NOT(relop(x,y)) ==> Reverse(relop)(x,y)
+                //
+                else if (VNFuncIsComparison(funcApp.m_func))
+                {
+                    *resultVN = GetRelatedRelop(arg0VN, VN_RELATION_KIND::VRK_Reverse);
+                }


This rewrite treats GT_NOT(relop) as logical negation of the comparison, but VN constant folding for GT_NOT is bitwise complement (~) and comparisons fold to 0/1. With 0/1 semantics, ~(x < y) is not equivalent to (x >= y), so this can make VNs unsound. Either represent logical NOT explicitly (e.g., XOR-with-1 / EQ-with-0) or gate this rewrite to a representation where true is all-bits-set and ensure folding matches it.

Suggested change

// NOT(relop(x,y)) ==> Reverse(relop)(x,y)

//

else if (VNFuncIsComparison(funcApp.m_func))

{

*resultVN = GetRelatedRelop(arg0VN, VN_RELATION_KIND::VRK_Reverse);

}

Copilot · 2026-04-20T21:05:46Z

+                    VNFuncApp arg1FN;
+                    if (GetVNFunc(arg1VN, &arg1FN) && VNFuncIsComparison(arg1FN.m_func))
+                    {
+                        if ((arg0FN.m_args[0] == arg1FN.m_args[0]) && (arg0FN.m_args[1] == arg1FN.m_args[1]))


This relop-combination simplification for GT_OR currently triggers for floating-point comparisons too, but several table identities are not valid with NaN/unordered semantics (e.g., LT | GE is not always true for floats). Consider skipping this optimization when either relop operand is floating-point (similar to GetRelatedRelop, which returns NoVN for float compares).

Suggested change

if ((arg0FN.m_args[0] == arg1FN.m_args[0]) && (arg0FN.m_args[1] == arg1FN.m_args[1]))

if ((arg0FN.m_args[0] == arg1FN.m_args[0]) && (arg0FN.m_args[1] == arg1FN.m_args[1]) &&

!varTypeIsFloating(TypeOfVN(arg0FN.m_args[0])) && !varTypeIsFloating(TypeOfVN(arg0FN.m_args[1])))

Copilot · 2026-04-20T21:05:46Z

+                VNFuncApp arg0FN;
+                if (GetVNFunc(arg0VN, &arg0FN) && VNFuncIsComparison(arg0FN.m_func))
+                {
+                    VNFuncApp arg1FN;
+                    if (GetVNFunc(arg1VN, &arg1FN) && VNFuncIsComparison(arg1FN.m_func))


Same issue for the GT_AND relop-combination: for floating-point comparisons, NaN/unordered semantics make several AND identities invalid. Add a non-floating check before consulting s_relatedRelopTable_AND.

Copilot · 2026-04-20T21:05:46Z

+                }
+                else if (andApp.m_args[0] == pathApp.m_args[1] && andApp.m_args[1] == pathApp.m_args[0])
+                {
+                    vnStore->GetRelatedRelop(andVN, ValueNumStore::VN_RELATION_KIND::VRK_Swap);


In the swapped-operands case, the return value of GetRelatedRelop(andVN, VRK_Swap) is ignored. GetRelatedRelop returns a new VN and does not mutate andVN, so andApp will still reflect the original (unswapped) relop. Assign the returned VN back to a local (e.g., andVN = ...) before re-querying andApp.

Suggested change

vnStore->GetRelatedRelop(andVN, ValueNumStore::VN_RELATION_KIND::VRK_Swap);

andVN = vnStore->GetRelatedRelop(andVN, ValueNumStore::VN_RELATION_KIND::VRK_Swap);

kg · 2026-04-20T22:26:27Z

        rii.domCmpNormVN = blockPathVN;

        optRelopImpliesRelop(&rii);
+        bool canOptimize = rii.canInfer && rii.canInferFromTrue && !rii.reverseSense;


optRelopImpliesRelop appears to exclude floating point values, so canOptimize will be false for FP operations, right? So copilot's stuff about FP is wrong?

Not quite, because we now do this new thing if that check fails, and that was not sufficiently cautious around FP compares.

kg · 2026-04-20T22:45:38Z

I went over this and don't feel confident enough in my understanding of VN to give it a green check. A lot of copilot's concerns match mine.

AndyAyersMS · 2026-04-21T15:19:09Z

Looks like there are some failures to sort through.

Copilot AI review requested due to automatic review settings April 20, 2026 20:54

github-actions Bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Apr 20, 2026

dotnet-policy-service Bot assigned AndyAyersMS Apr 20, 2026

Copilot started reviewing on behalf of AndyAyersMS April 20, 2026 20:56 View session

Copilot AI reviewed Apr 20, 2026

View reviewed changes

kg reviewed Apr 20, 2026

View reviewed changes

build-analysis Bot mentioned this pull request Apr 21, 2026

Build error: ilc exited with code 57005 #124976

Open

review feedback

e4c89c9

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

JIT: relop simplification during redundant branch opts#127181

JIT: relop simplification during redundant branch opts#127181
AndyAyersMS wants to merge 2 commits intodotnet:mainfrom
AndyAyersMS:RboSimplifyForDominatingRelop

AndyAyersMS commented Apr 20, 2026

Uh oh!

AndyAyersMS commented Apr 20, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 20, 2026

Uh oh!

Copilot AI Apr 20, 2026

Uh oh!

Copilot AI Apr 20, 2026

Uh oh!

Copilot AI Apr 20, 2026

Uh oh!

Copilot AI Apr 20, 2026

Uh oh!

Copilot AI Apr 20, 2026

Uh oh!

kg Apr 20, 2026 •

edited

Loading

Uh oh!

AndyAyersMS Apr 21, 2026

Uh oh!

kg commented Apr 20, 2026

Uh oh!

AndyAyersMS commented Apr 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

	if ((arg0FN.m_args[0] == arg1FN.m_args[0]) && (arg0FN.m_args[1] == arg1FN.m_args[1]))
	if ((arg0FN.m_args[0] == arg1FN.m_args[0]) && (arg0FN.m_args[1] == arg1FN.m_args[1]) &&
	!varTypeIsFloating(TypeOfVN(arg0FN.m_args[0])) && !varTypeIsFloating(TypeOfVN(arg0FN.m_args[1])))

	vnStore->GetRelatedRelop(andVN, ValueNumStore::VN_RELATION_KIND::VRK_Swap);
	andVN = vnStore->GetRelatedRelop(andVN, ValueNumStore::VN_RELATION_KIND::VRK_Swap);

Conversation

AndyAyersMS commented Apr 20, 2026

Uh oh!

AndyAyersMS commented Apr 20, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

kg Apr 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

AndyAyersMS Apr 21, 2026

Choose a reason for hiding this comment

Uh oh!

kg commented Apr 20, 2026

Uh oh!

AndyAyersMS commented Apr 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

kg Apr 20, 2026 •

edited

Loading